Traffic large model enabled micro base station power dynamic regulation and control method and related equipment
By combining traffic big data language models with reinforcement learning, future traffic prediction and dynamic power regulation strategies are generated, solving the problems of lag and energy waste in micro base station regulation strategies, and realizing forward-looking prediction of traffic changes and optimization of network performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-21
- Publication Date
- 2026-04-14
AI Technical Summary
Existing micro base station control strategies are difficult to dynamically and adaptively adjust based on current urban traffic flow, resulting in energy waste and network response lag. Traditional traffic prediction models are insufficiently applied in the field of communication networks and lack forward-looking control capabilities.
Traffic flow data is predicted using a traffic big data model to generate future flow data, which is then integrated with the current network state to form a multi-dimensional joint state space. Dynamic power regulation strategies are generated through reinforcement learning, and the power adjustment of micro base stations is optimized using the ϵ-greedy algorithm and reward function.
It enables proactive prediction of traffic change trends, supports advance power adjustments, reduces redundant energy consumption of micro base stations, and improves network performance and user experience.
Smart Images
Figure CN121865384A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of dynamic power control technology for micro base stations, specifically to a method and related equipment for dynamic power control of micro base stations powered by a large traffic model. Background Technology
[0002] 5G networks are accelerating the construction of smart cities and intelligent transportation systems. Facing multiple challenges such as high-density user access, a surge in ultra-low latency services, and green and low-carbon development, typical application scenarios (such as vehicle-to-everything (V2X) collaborative sensing, communication support for large-scale events, and emergency command and dispatch) place higher demands on the dynamic adaptability of network resources. In complex urban scenarios (such as transportation hubs, commercial centers, and areas surrounding large event venues), user distribution exhibits significant spatiotemporal characteristics, resulting in hotspot areas. Hotspot areas typically refer to scenarios with a higher density of users. Micro base stations, with their advantages of flexible deployment and precise coverage, have become a key infrastructure for improving network capacity and energy efficiency, and are commonly used to cover these hotspot areas. However, due to the cyclical nature of traffic in these scenarios (such as train stations and shopping malls), the number of users does not always remain at a high density, leading to energy waste during off-peak hours.
[0003] Currently, mainstream micro base station control strategies still rely on historical statistics and static rules (such as timed triggering of start-stop power regulation). While simple to implement and highly stable, these strategies struggle to adapt to the non-steady-state demands brought about by the spatiotemporal dynamics of traffic and the suddenness of events. Furthermore, with the development of artificial intelligence, machine learning is increasingly being applied to intelligent transportation systems. Reinforcement learning (RL) is a machine learning method that optimizes sequential decision-making strategies based on reward signals through continuous interaction between an agent and its environment. It possesses adaptive dynamic environment and online learning capabilities, enabling it to efficiently solve complex decision-making problems. To enhance the intelligence of control, some research attempts to introduce pure reinforcement learning, using real-time network load and other indicators as state inputs to directly learn micro base station start-stop or power adjustment strategies.
[0004] While such methods possess optimization potential, they still have significant limitations. They lack forward-looking control capabilities and are essentially "reactive" control, meaning they only start / stop or adjust power based on current or recent network load, lacking the ability to predict traffic change trends. This lag leads to untimely base station responses, easily causing localized network congestion; and during low-load periods, they continue to operate because they cannot detect traffic decline in advance, resulting in significant energy waste.
[0005] Although traffic flow prediction technology has been widely applied in the field of urban traffic management (such as signal control optimization and congestion warning), its prediction targets are usually traffic density or traffic speed. There is little research on migrating its results to the field of communication networks to predict the future potential number of mobile user accesses in a specific geographical area. This is mainly because traditional traffic prediction models generally adopt lightweight small models (such as ARIMA, LSTM), whose modeling capabilities are limited and it is difficult to capture the complex long-term temporal dependencies in urban traffic flow, resulting in insufficient prediction accuracy and generalization ability, and being unable to support real-time and highly reliable micro base station regulation decisions.
[0006] These problems have caused the micro base station to not be able to dynamically control the power in advance according to the number of users, restricting the forward-looking regulation of the power of 5G micro base stations. Summary of the Invention
[0007] Based on the problems raised in the above background technology, the purpose of the present invention is to provide a method for dynamically regulating the power of a micro base station empowered by a traffic large model and related devices, which solves the problem that it is currently difficult for a micro base station to dynamically and adaptively adjust the transmission power according to the current urban traffic flow.
[0008] The present invention is realized through the following technical solutions:
[0009] The first aspect of the present invention provides a method for dynamically regulating the power of a micro base station empowered by a traffic large model, including: future traffic prediction and power regulation;
[0010] Among them, the future traffic prediction includes the following steps:
[0011] Preprocess and transform the obtained traffic flow data to generate a time-series spatial text data set;
[0012] Input the time-series spatial text data set into the fine-tuned large language model to obtain the future traffic prediction value;
[0013] Among them, the power regulation includes the following steps:
[0014] Observe the current real-time network state and the current traffic flow situation in the regulation area, and jointly encode the future traffic prediction value with the current real-time network state and the current traffic flow situation to generate a multi-dimensional joint state space;
[0015] Construct an action space and an agent, and the agent observes the multi-dimensional joint state space and selects a joint power action from the action space using the ϵ-greedy algorithm;
[0016] Obtain the network performance index after the agent executes the joint power action, calculate the reward function according to the network performance index, and obtain the multi-dimensional joint state space at the next moment;
[0017] Repeated power regulation enables iterative learning of the agent and regulation regions, outputting a dynamic power regulation strategy for micro base stations.
[0018] The aforementioned technical solution comprises two parts: future traffic flow prediction and power regulation. In future traffic flow prediction, a large language model is used to predict the future traffic flow data of the area to be predicted, thus enabling the prediction of future traffic flow in urban hotspots. By transferring traffic flow prediction to the field of communication networks to predict the number of potential mobile users accessing a specific geographical area, a forward-looking prediction of traffic change trends is achieved.
[0019] For the power regulation part, the regulation area is constructed as the environment. The predicted traffic data of the area to be predicted in the next unit time output by the trained large language model is spliced and fused with the current real-time observation data of the micro base station to form a multi-dimensional joint state space, realizing the joint perception of "future trend + current state". The execution action of micro base station power control is constructed as the action space. The reward function is calculated based on the execution of joint power action. The current state space, action space, reward function and the state space of the next round are integrated into a quadruple and stored in the replay buffer for training the agent. Through the continuous interaction and online learning between the agent and the environment, a dynamic and adaptive power regulation strategy is generated online, realizing the collaborative optimization goal of significantly reducing the redundant energy consumption of micro base stations while ensuring the performance of user network.
[0020] In one optional embodiment, the acquired traffic flow data is preprocessed and transformed, including the following steps:
[0021] The traffic flow data is preprocessed to obtain a structured data table; wherein the horizontal axis of the structured data table is the street identifier and the vertical axis is the time series.
[0022] The structured data table is converted into a structured text format recognizable by a large language model; wherein the structured text format includes:
[0023] Task Instructions: Based on the locational relationships of each street and the historical traffic flow of each street, predict the traffic flow in the designated area;
[0024] Input: Historical traffic flow data for each street / unit over a given time period;
[0025] Output: Traffic flow data for the specified area.
[0026] In one alternative embodiment, the fine-tuning process of the large language model includes:
[0027] Freeze all original parameters in the large language model and inject a trainable low-rank adapter into the projection matrix of Query and Value in the self-attention mechanism of each Transformer layer.
[0028] Use the AdamW optimizer to select parameters for training configuration during training;
[0029] After training, the original weights of the large language model and the LoRA weights are merged and overridden by reparameterization to generate a fine-tuned large language model.
[0030] In one optional embodiment, the current real-time network status and current traffic flow include: the number of users accessing the i-th micro base station at time t, the network performance of user communication on the current map, and the transmission power of the i-th micro base station at time t;
[0031] The predicted future traffic values and the current real-time network state are jointly encoded to generate a multi-dimensional joint state space, including:
[0032] The predicted future traffic is spliced and fused with the number of users accessing the i-th micro base station at time t, the network performance of user communication on the current map, and the transmission power of the i-th micro base station at time t to generate a multi-dimensional joint state space.
[0033] In one alternative embodiment, constructing the action space includes:
[0034] Define a set of selectable transmit power levels for each micro base station; wherein, the transmit power levels in the set of selectable transmit power levels are arranged at equal intervals with a fixed step size;
[0035] Each micro base station independently executes actions based on a set of selectable transmit power levels, resulting in the actions of each micro base station.
[0036] The actions of each micro base station are combined using a Cartesian product to generate an action space.
[0037] In one optional embodiment, the ϵ-greedy algorithm is used to select joint power actions from the action space, including: the agent selects each micro base station transmit power combination that has not been selected in the action space with probability ϵ; and selects the optimal transmit power combination in the action space with probability 1-ϵ.
[0038] In one optional embodiment, obtaining network performance metrics after the agent performs the joint power action, and calculating a reward function based on the network performance metrics, includes:
[0039] The first dynamic configuration item is constructed using the aforementioned network performance metrics;
[0040] A second dynamic configuration item is constructed based on the transmit power of each micro base station;
[0041] The reward function is obtained by weighting the first dynamic configuration item and the second dynamic configuration item.
[0042] In one optional embodiment, repetitive power modulation enables iterative learning of the agent and modulation regions, including:
[0043] The multidimensional joint state space, joint power action, reward function at the current moment, and the multidimensional joint state space at the next moment are integrated into experience and stored in the experience replay buffer pool;
[0044] The agent is trained using the experience in the experience replay buffer pool;
[0045] The interactive training between the control region and the agent in repetitive power regulation ultimately outputs a dynamic power regulation strategy for micro base stations.
[0046] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for dynamic power control of micro base stations enabled by a traffic big model.
[0047] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for dynamic power control of micro base stations empowered by a large traffic model.
[0048] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0049] 1. This invention transfers traffic flow prediction to the field of communication networks, enabling forward-looking prediction of traffic flow change trends;
[0050] 2. This invention splices and fuses the future traffic data of the area to be predicted with the current real-time network status to construct a joint state space, enabling the agent to not only accurately identify whether there is a risk of congestion in the current network, but also to predict in advance whether the network will become congested in the short term, thereby supporting advance and smooth power increase or decrease operations.
[0051] 3. By weighting network performance and micro base station transmit power, a reward function is obtained, which avoids the strategy being overly biased towards energy saving, thus avoiding a deterioration in user experience, and balances the control accuracy and feasibility. Attached Figure Description
[0052] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:
[0053] Figure 1 This is a schematic diagram illustrating the fine-tuning of a pre-trained large language model in the prediction section provided in Embodiment 1 of the present invention;
[0054] Figure 2 This is a schematic diagram of reinforcement learning in the regulation part provided in Embodiment 1 of the present invention;
[0055] Figure 3 This is a schematic diagram of the structure of an electronic device provided in Embodiment 2 of the present invention. Detailed Implementation
[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.
[0057] Example 1
[0058] Embodiment 1 of the present invention provides a method for dynamic power control of micro base stations empowered by a large traffic model. The core of this method includes two parts: prediction and control.
[0059] In this embodiment, the prediction part uses a large language model to predict the processed traffic flow data, thereby obtaining the future traffic flow data for the area to be predicted, thus realizing the prediction of future traffic flow in urban hotspot areas, such as... Figure 1 As shown, the processed traffic flow data is predicted using a large language model, including:
[0060] Traffic flow data is acquired and preprocessed using Python scripts to remove outliers and zero values (e.g., invalid records due to sensor malfunctions), thereby reducing interference and obtaining a structured data table (i.e., the data is reshaped through preprocessing). Figure 1 The leftmost example is the spatiotemporal heatmap. The structured data table is a two-dimensional matrix with street identifiers on the horizontal axis and time series on the vertical axis, thus forming a "street-time" two-dimensional flow matrix.
[0061] Then, the structured data table is divided according to the spatiotemporal dimension, and the location information of different regions is integrated to indicate the relationships between streets (e.g., Figure 1As shown in ST-Data, this enhances the correlation between regional locations, thereby improving the ability of subsequent models to capture these correlations.
[0062] The preprocessed data is converted into a structured text format that can be recognized by a large language model. This structured text format (such as...) Figure 1 The instruction (as shown in the text) consists of three parts:
[0063] Instruction: Clearly define the task instructions, such as: "Based on all street signs and historical traffic data, combined with regional location relationships, predict the traffic flow of the target area within a unit of time in the future."
[0064] Input: Contains traffic flow for each street unit over a given time period;
[0065] Output: Traffic flow in the area to be predicted.
[0066] The instruction, input, and output are combined to form a temporal and spatiotemporal text dataset for training large language models.
[0067] Then, the temporal spatial text dataset with explicit task instructions is input into the pre-trained Large Language Model (LLM) for fine-tuning. In this embodiment, the LoRA fine-tuning method is used: all original parameters in the Large Language Model are frozen; in the self-attention mechanism of each Transformer layer, a trainable low-rank adapter is injected by wayside into the projection matrix of Query and Value; the AdamW optimizer is used to select appropriate learning rate, weight decay and other parameters for training configuration during training; after training, the original weights of the pre-trained Large Language Model and the LoRA weights are merged and overlaid on the original module weights through reparameterization to generate a Large Language Model for traffic flow prediction with a structure completely consistent with the original model and lossless inference efficiency.
[0068] By fine-tuning the large language model, the final output is the predicted traffic data for the area to be predicted in the next unit of time for that map unit.
[0069] In this embodiment, the regulation part constructs the regulation area as the environment, in which one macro base station and M micro base stations are deployed. The predicted traffic data of the area to be predicted in the next unit time output by the fine-tuned large language model is spliced and fused with the current real-time observation data of the micro base stations to form a multi-dimensional joint state space. The execution actions of micro base station power control are constructed as the action space. The reward function is calculated based on the execution of the joint power actions. The current state space, action space, reward function, and the state space of the next round are integrated into a four-tuple and stored in the replay buffer for training the agent. Through continuous interaction and online learning between the agent and the environment, a dynamic and adaptive power regulation strategy is generated online, achieving the collaborative optimization goal of significantly reducing the redundant energy consumption of micro base stations while ensuring user network performance. Specifically, the regulation stage is as follows: Figure 2 As shown, it includes:
[0070] Step S1: Within the control area, deploy 1 macro base station and M micro base stations to construct the control area as the environment; obtain the current real-time network status and current traffic flow by observing the environment. In this embodiment, the current real-time network status includes: the network performance of user communication on the current map (e.g., total packet loss rate). Average end-to-end delay (throughput, etc.), the transmit power P of the i-th micro base station at time t i，t (i=1,2,…,M); Current traffic flow includes: N number of users connected to the i-th micro base station at time t. i，t (i=1,2,…,M).
[0071] The current real-time network status and traffic flow are compared with the predicted traffic data for the area to be predicted in the next unit of time, output by the large language model. The joint encoding constitutes the joint state space at time t. Specifically, the joint state space at time t It is expressed as follows:
[0072]
[0073] By splicing and fusing the current real-time network status, current traffic flow, and predicted traffic data of the area to be predicted, a multi-dimensional joint state space of the reinforcement learning controller is constructed, realizing the joint perception of "future trends + current status". This enables the agent to not only accurately identify whether there is a risk of congestion in the current network, but also to predict in advance whether the network will become congested in the short term, thereby supporting early and smooth power increase or decrease operations.
[0074] Step S2: Construct the action space for 5G micro base station power control. In actual deployment, this is achieved through the collaborative coverage of M micro base stations. Specifically, constructing the action space includes the following steps:
[0075] First, define the set of selectable transmit power levels for each micro base station: For the i-th micro base station (i=1,2,…,M), its set of selectable transmit power levels is defined as:
[0076]
[0077] Among them, 0, , These represent the current power levels of the micro base station. This indicates the maximum allowed transmission power of the micro base station. For example, 0 indicates that the current power of the micro base station is 0, and selecting 0 means turning off the micro base station.
[0078] Furthermore, They are arranged at fixed intervals with equal step sizes. The power determines the coverage area of the micro base station; the greater the power, the greater the coverage area.
[0079] Secondly, the action space is constructed based on the set of selectable transmit power levels for each micro base station:
[0080] Each micro base station executes its actions independently, that is, it selects a transmission power level from the set of available transmission power levels for the micro base station. The reason for this is that the traffic distribution in the areas served by each micro base station has spatial heterogeneity, and their optimal control power is also different. Therefore, each micro base station will execute its actions independently, resulting in the actions of each micro base station.
[0081] The actions of each microbase station are combined using a Cartesian product to generate the action space at time t. It is represented as follows:
[0082]
[0083] The action space is defined as a set of discretized power levels, that is, the action space represents the set of transmit power selected by each micro base station.
[0084] Step S3: Construct a proxy. The proxy observes the joint state space at time t. And the ϵ-greedy algorithm is used in the action space at time t. Choose an action.
[0085] Specifically, the agent randomly explores untried power combinations with probability ϵ and selects the optimal action output by the current policy network with probability 1−ϵ, which is the optimal transmit power combination selected by each micro base station, to ensure sufficient exploration in the early stage of training and convergence to a stable policy in the later stage.
[0086] The agent performs the above action, that is, adjusts the transmit power of each micro base station to the action space at time t. The transmit power selected by each micro base station is then used to output the optimal power level index, i.e., the power configuration command.
[0087] Step S4: Obtain the current network performance metrics after the action is executed, such as packet loss rate and average latency, and calculate the reward function at time t using the current network performance metrics. This reward function balances network quality, taking into account both control precision and feasibility, and avoids the strategy from being overly biased towards energy saving, which could lead to a deterioration in user experience.
[0088] Furthermore, a multi-objective weighted reward function is designed. The construction of this reward function includes: using network performance indicators as the first dynamic configuration item, using the transmit power of micro base stations as the second dynamic configuration item, and constructing the reward function through weight coefficients.
[0089] Taking the average packet loss rate as an example, the first dynamic configuration item is constructed using the average packet loss rate of the entire network or key areas at the current moment, thus reflecting communication reliability; the second dynamic configuration item is constructed using the ratio of the sum of the current transmit power of all micro base stations to the sum of the maximum power of all micro base stations. Specifically, the reward function is expressed as follows:
[0090]
[0091] in, This represents the average packet loss rate of the entire network or key areas at time t. This is the sum of the current transmit power of all micro base stations. The sum of the maximum power of all micro base stations is normalized and used as the energy consumption cost item, with a weighting coefficient. It can be dynamically configured according to needs (such as upgrading for high-priority business scenarios). Energy-saving mode improvement ).
[0092] After obtaining the current network performance metrics, substitute them into the reward function to calculate the reward function at time t. And re-observe the joint space state at the next moment. .
[0093] Step S5: Determine the joint space state at time t. Action space Reward function Joint space state at time t+1 Integrate into four data sets {( , , , The data is stored in the ExperienceReplay Buffer. The agent is trained using the experience (four-data set) in the ExperienceReplay Buffer to complete the batch training and parameter update of the network. Steps S1 to S5 are repeated to realize the continuous interaction and online learning between the agent and the environment. Through the iterative output of the agent and the environment, the micro base station power dynamic control strategy is output, which significantly improves the base station's adaptive capability and realizes the forward-looking control of 5G micro base station power, providing more reliable support for the dynamic decision-making of intelligent transportation systems.
[0094] Example 2
[0095] Figure 3 This is a schematic diagram of the structure of an electronic device provided in Embodiment 2 of the present invention, as shown below. Figure 3 As shown, the electronic device includes a processor 21, a memory 22, an input device 23, and an output device 24; the number of processors 21 in the computer device can be one or more. Figure 3 Taking a processor 21 as an example; the processor 21, memory 22, input device 23, and output device 24 in an electronic device can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.
[0096] The memory 22, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor 21 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 22, thereby realizing the dynamic power control method for micro base stations empowered by the traffic large model in Embodiment 1.
[0097] The memory 22 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 22 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include memory remotely located relative to the processor 21, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
[0098] Input device 23 can be used to receive user input such as ID and password. Output device 24 is used to output the network configuration page.
[0099] Example 3
[0100] Embodiment 3 of the present invention also provides a computer-readable storage medium, wherein the computer-executable instructions, when executed by a computer processor, are used to implement the micro base station power dynamic control method empowered by the traffic large model as provided in Embodiment 1.
[0101] The storage medium containing computer-executable instructions provided in the embodiments of the present invention is not limited to the method operation provided in Embodiment 1, but can also execute related operations in the traffic large model-enabled micro base station power dynamic control method provided in any embodiment of the present invention.
[0102] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for dynamic power control of micro base stations empowered by a large traffic model, characterized in that, include: Future flow forecasting and power regulation; Future traffic forecasting includes the following steps: The acquired traffic flow data is preprocessed and transformed to generate a time-series spatial text dataset; The temporal spatial text dataset is input into the fine-tuned large language model to obtain future traffic prediction values; Power regulation includes the following steps: The current real-time network status and current traffic flow in the observation and control area are used to jointly encode the future traffic flow prediction value and the current real-time network status and current traffic flow to generate a multi-dimensional joint state space. An action space and an agent are constructed. The agent selects joint power actions from the action space by observing the multidimensional joint state space using the ϵ-greedy algorithm. Obtain the network performance metrics after the agent performs the joint power action, calculate the reward function based on the network performance metrics, and obtain the multidimensional joint state space at the next time step; Repeated power regulation enables iterative learning of the agent and regulation regions, outputting a dynamic power regulation strategy for micro base stations.
2. The method for dynamic power control of micro base stations empowered by a large traffic model according to claim 1, characterized in that, The acquired traffic flow data undergoes preprocessing and data transformation, including the following steps: The traffic flow data is preprocessed to obtain a structured data table; wherein the horizontal axis of the structured data table is the street identifier and the vertical axis is the time series. The structured data table is converted into a structured text format recognizable by a large language model; wherein the structured text format includes: Task Instructions: Based on the locational relationships of each street and the historical traffic flow of each street, predict the traffic flow in the designated area; Input: Historical traffic flow data for each street / unit over a given time period; Output: Traffic flow data for the specified area.
3. The method for dynamic power control of micro base stations empowered by a large traffic model according to claim 1, characterized in that, The fine-tuning process of a large language model includes: Freeze all original parameters in the large language model and inject a trainable low-rank adapter into the projection matrix of Query and Value in the self-attention mechanism of each Transformer layer. Use the AdamW optimizer to select parameters for training configuration during training; After training, the original weights of the large language model and the LoRA weights are merged and overridden by reparameterization to generate a fine-tuned large language model.
4. The method for dynamic power control of micro base stations empowered by a large traffic model according to claim 1, characterized in that, The current real-time network status and current traffic flow include: the number of users connected to the i-th micro base station at time t, the network performance of user communication on the current map, and the transmission power of the i-th micro base station at time t; The predicted future traffic values and the current real-time network state are jointly encoded to generate a multi-dimensional joint state space, including: The predicted future traffic is spliced and fused with the number of users accessing the i-th micro base station at time t, the network performance of user communication on the current map, and the transmission power of the i-th micro base station at time t to generate a multi-dimensional joint state space.
5. The method for dynamic power control of micro base stations empowered by a large traffic model according to claim 4, characterized in that, Constructing the action space includes: Define a set of selectable transmit power levels for each micro base station; wherein, the transmit power levels in the set of selectable transmit power levels are arranged at equal intervals with a fixed step size; Each micro base station independently executes actions based on a set of selectable transmit power levels, resulting in the actions of each micro base station. The actions of each micro base station are combined using a Cartesian product to generate an action space.
6. The method for dynamic power control of micro base stations empowered by a large traffic model according to claim 5, characterized in that, The ϵ-greedy algorithm is used to select joint power actions from the action space, including: the agent selects each micro base station transmit power combination that has not been selected in the action space with probability ϵ; and selects the optimal transmit power combination in the action space with probability 1-ϵ.
7. The method for dynamic power control of micro base stations empowered by a large traffic model according to claim 1, characterized in that, Obtain the network performance metrics after the agent performs the joint power action, and calculate the reward function based on the network performance metrics, including: The first dynamic configuration item is constructed using the aforementioned network performance metrics; A second dynamic configuration item is constructed based on the transmit power of each micro base station; The reward function is obtained by weighting the first dynamic configuration item and the second dynamic configuration item.
8. The method for dynamic power control of micro base stations empowered by a large traffic model according to claim 1, characterized in that, Repetitive power regulation enables iterative learning of the agent and regulation region, including: The multidimensional joint state space, joint power action, reward function at the current moment, and the multidimensional joint state space at the next moment are integrated into experience and stored in the experience replay buffer pool; The agent is trained using the experience in the experience replay buffer pool; The interactive training between the control region and the agent in repetitive power regulation ultimately outputs a dynamic power regulation strategy for micro base stations.
9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for dynamic power control of micro base stations empowered by a traffic large model as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for dynamic power control of micro base stations empowered by a large traffic model as described in any one of claims 1 to 8.